When a Quantum Processor and a Classical Computer Work Together

The research centred on variational quantum algorithms and quantum generative machine learning — methods in which a quantum processor and a classical computer work in tandem. The quantum system generates quantum states or samples, whilst the classical algorithm optimises the model parameters. According to the researchers, these approaches are particularly promising for the near-term, still limited quantum computers of the Noisy Intermediate-Scale Quantum (NISQ) era.

The team developed several new algorithms and substantially improved their efficiency. Among their achievements, they developed techniques that simplify the construction of quantum circuits, allowing the same computations to be carried out with fewer quantum gates. This is of paramount importance, since current quantum computers are extremely sensitive to noise and errors, making every gate operation count.

One of the project's most significant results was the development of a new optimisation strategy. A fundamental challenge in training variational quantum algorithms is that the optimiser frequently becomes trapped: changes in the system become so minimal that the learning process effectively stalls. The researchers developed a new method capable of avoiding these flat optimisation regions. Rather than adjusting parameters in small incremental steps, the approach makes larger, directed jumps through the solution space. This optimisation procedure was successfully tested on large-scale quantum circuits, proving effective even for quantum circuits containing tens of thousands of entangling quantum gates. The researchers further enhanced the algorithm with elements based on evolutionary selection, which improved its ability to avoid unfavourable local minima and more efficiently locate globally optimal solutions.

A second key focus of the research was the study of quantum generative models. The team developed models that, owing to their special structure, can be trained entirely on classical hardware, while efficient sampling remains possible only on a quantum computer. In this context, two new models were introduced: the fermionic and bosonic Born machines, which employ parameterisable fermionic and bosonic linear optical transformations respectively. The researchers demonstrated excellent classical training scalability, successfully handling photonic systems with up to 805 modes and with over one million trainable parameters.

The project's findings make a fundamental contribution to our understanding of how classical and quantum computational methods might work together in the intelligent systems of the future. In the longer term, the research may open new avenues for addressing challenging optimisation problems, advancing artificial intelligence, performing quantum chemistry simulations, and processing highly complex datasets.